Modern deep neural networks rely on depth and model complexity to enhance their representational power, yet excessive network depth introduces a critical challenge: supervision vanishing in intermediate layers. In this paper, we propose Locality-aware Backward Propagation (LBP), a training framework that resolves this through a dual-phase mechanism: in Stage 1, strict gradient confinement ensures segment-specific subgraph optimization without interference from distant gradients; in Stage 2, global coherence is harmonized via multi-loss-driven updates. By decoupling localized supervision from global objectives, LBP not only alleviates supervision vanishing but also enables stable gradient accumulation, where global attention mechanisms amplify parameter sensitivity. The experimental results demonstrate that our proposed method not only achieves competitive performance on CNN architectures across both CIFAR-10/100 datasets but also delivers remarkable improvements on Vision Transformers.